!9870 modify some cpu kernel thread num

From: @kisnwang
Reviewed-by: @zhoufeng54,@jjfeing,@chujinjin
Signed-off-by: @chujinjin

mindspore/ccsrc/backend/kernel_compiler/cpu/adam_delta_cpu_kernel.cc

@@ -24,7 +24,11 @@
 namespace mindspore {
 namespace kernel {
 constexpr size_t kAdamDeltaInputSize = 9;
+#ifdef ENABLE_D
 constexpr size_t kUsedThreadNum = 23;
+#else
+constexpr size_t kUsedThreadNum = 8;
+#endif
 namespace {
 struct ComputeParam {
   float *delta_{nullptr};

mindspore/ccsrc/backend/kernel_compiler/cpu/embedding_look_up_cpu_kernel.cc

@@ -22,6 +22,11 @@
 namespace mindspore {
 namespace kernel {
 namespace {
+#ifdef ENABLE_D
+constexpr size_t kUsedThreadNum = 23;
+#else
+constexpr size_t kUsedThreadNum = 8;
+#endif
 template <typename T>
 void LookUpTableTask(const float *input_addr, const T *indices_addr, float *output_addr, size_t indices_lens,
                      size_t outer_dim_size, T offset, size_t first_dim_size) {
@@ -92,10 +97,9 @@ void EmbeddingLookUpCPUKernel::LaunchKernel(const std::vector<kernel::AddressPtr
   auto input_addr = reinterpret_cast<float *>(inputs[0]->addr);
   auto indices_addr = reinterpret_cast<T *>(inputs[1]->addr);
   auto output_addr = reinterpret_cast<float *>(outputs[0]->addr);
-  const size_t kMaxThreadNum = 16;
   size_t thread_num = indices_lens_ / 10000 + 1;
-  thread_num = thread_num > kMaxThreadNum ? kMaxThreadNum : thread_num;
-  std::thread threads[kMaxThreadNum];
+  thread_num = thread_num > kUsedThreadNum ? kUsedThreadNum : thread_num;
+  std::thread threads[kUsedThreadNum];
   size_t task_proc_lens = (indices_lens_ + thread_num - 1) / thread_num;
   size_t i;
   size_t task_offset = 0;

mindspore/ccsrc/backend/kernel_compiler/cpu/scatter_nd_update_cpu_kernel.cc

@@ -22,6 +22,11 @@
 namespace mindspore {
 namespace kernel {
 namespace {
+#ifdef ENABLE_D
+constexpr size_t kUsedThreadNum = 23;
+#else
+constexpr size_t kUsedThreadNum = 8;
+#endif
 template <typename T>
 void Compute(const ComputeParams<T> *params, const size_t start, const size_t end) {
   MS_EXCEPTION_IF_NULL(params);
@@ -115,10 +120,9 @@ void ScatterNdUpdateCPUKernel::LaunchKernel(const std::vector<AddressPtr> &input
   params.indices_unit_rank_ = indices_unit_rank_;
   params.out_strides_ = &out_strides_;
 
-  const size_t thread_num = 24;
   std::vector<Task> tasks;
   size_t start = 0;
-  size_t once_compute_size = (num_units_ + thread_num - 1) / thread_num;
+  size_t once_compute_size = (num_units_ + kUsedThreadNum - 1) / kUsedThreadNum;
   while (start < num_units_) {
     size_t end = (start + once_compute_size) > num_units_ ? num_units_ : (start + once_compute_size);
     auto task = [&params, start, end]() -> int {

mindspore/ccsrc/backend/kernel_compiler/cpu/sparse_optimizer_cpu_kernel.h

@@ -27,6 +27,11 @@
 
 namespace mindspore {
 namespace kernel {
+#ifdef ENABLE_D
+constexpr size_t kUsedThreadNum = 23;
+#else
+constexpr size_t kUsedThreadNum = 8;
+#endif
 template <typename T>
 struct SparseGradient {
   float *value_{nullptr};
@@ -95,7 +100,7 @@ class SparseOptimizerCPUKernel : public CPUKernel {
   static void BucketReduceSparseGradient(const ReduceSparseGradientParam<T> &param) {
     MS_LOG(DEBUG) << "Start";
     MS_EXCEPTION_IF_NULL(param.input_grad_);
-    size_t thread_num = 23;
+    size_t thread_num = kUsedThreadNum;
     if (param.input_grad_->indices_size_ < thread_num) {
       thread_num = param.input_grad_->indices_size_;
     }
@@ -120,11 +125,10 @@ class SparseOptimizerCPUKernel : public CPUKernel {
   template <typename T>
   void MultiThreadCompute(const MultiThreadComputeFunc<T> &func, MultiThreadComputeParams<T> *params,
                           size_t total_compute_size) const {
-    const size_t kThreadNum = 24;
     std::vector<std::thread> threads;
-    threads.reserve(kThreadNum);
+    threads.reserve(kUsedThreadNum);
     size_t start = 0;
-    size_t once_compute_size = (total_compute_size + kThreadNum - 1) / kThreadNum;
+    size_t once_compute_size = (total_compute_size + kUsedThreadNum - 1) / kUsedThreadNum;
     while (start < total_compute_size) {
       size_t end = (start + once_compute_size) > total_compute_size ? total_compute_size : (start + once_compute_size);
       threads.emplace_back(std::thread(func, params, start, end));

mindspore/ccsrc/backend/kernel_compiler/cpu/unique_cpu_kernel.cc

@@ -20,7 +20,11 @@
 namespace mindspore {
 namespace kernel {
 const size_t kUseBucketUniqueSize = 100000;
-const size_t kUniqueThreadNum = 23;
+#ifdef ENABLE_D
+constexpr size_t kUniqueThreadNum = 23;
+#else
+constexpr size_t kUniqueThreadNum = 8;
+#endif
 void UniqueCPUKernel::InitKernel(const CNodePtr &kernel_node) {
   node_ = kernel_node;
   CheckParam(kernel_node);